RealDeal: Enhancing Realism and Details in Brain Image Generation via Image-to-Image Diffusion Models.

Area of Science:

• Biomedical imaging
• Artificial intelligence
• Medical image generation

Background:

• Generative models, particularly latent diffusion models (LDMs), excel at generating brain MRIs.
• However, LDM-generated images often lack realism due to over-smoothing and missing fine anatomical details and noise.
• This limits their utility in applications requiring high fidelity.

Purpose of the Study:

• To enhance the realism and add fine details to brain MRI images generated by latent diffusion models.
• To introduce sharp edges, textures, anatomical features, and imaging noise characteristic of real scans.
• To formulate this enhancement process using an image-to-image diffusion model framework.

Main Methods:

• Developed an image-to-image diffusion model to refine LDM-generated brain MRI images.
• Introduced techniques to add sharp edges, fine textures, subtle anatomical features, and imaging noise.
• Utilized standard metrics (FID, LPIPS) for realism assessment.
• Developed novel metrics to quantify improvements in noise distribution, sharpness, and texture.

Main Results:

• The proposed image-to-image diffusion model successfully enhanced the realism of LDM-generated brain MRIs.
• Generated images exhibited improved sharpness, finer textures, and more realistic anatomical structures.
• The addition of subtle imaging noise increased the fidelity of the synthetic images.
• Novel metrics confirmed significant improvements in noise, sharpness, and texture.

Conclusions:

• Image-to-image diffusion models can effectively address the smoothness limitations of LDMs in brain MRI generation.
• The proposed method significantly improves the realism and detail of synthetic brain MRIs.
• This advancement holds potential for improving AI-driven applications in medical imaging analysis and simulation.